The grid used to be simple. Utilities built power plants, regulators set rates, customers paid bills. The price of electricity was a negotiated average, spread across millions of ratepayers, managed by commissions that moved at the speed of bureaucracy. That model is collapsing under the weight of a single new buyer class: AI data centers signing power purchase agreements so large they're rewriting how electricity gets priced, allocated, and paid for.
This is the story I've been tracking since PJM threatened to curtail data center loads in August. The threat was a symptom. The disease is structural: a demand curve that's growing faster than the pricing mechanisms designed to manage it.
The Stranded Cost Problem Nobody Wants to Own
Here's the core tension. When a hyperscaler signs a letter of intent for 500 MW of new load, utilities start building — substations, high-voltage lines, sometimes entire new generation commitments — years before the first server rack arrives. Some data centers get built on schedule. Some shrink. Some evaporate entirely. The steel goes in the ground on the promise. If the project changes, the cost doesn't.
That's the stranded cost problem. And it's not theoretical. Utilities in Ohio and Virginia are already grappling with it, which is why both states have moved to implement new tariff structures for large data center loads — requiring long-term commitments, payment for reserved capacity even if unused, and upfront collateral. The goal is to shift financial risk back onto the operators and away from residential ratepayers who had nothing to do with the investment decision.
This is the right instinct. The old model — where speculative infrastructure costs get socialized across the rate base — is a subsidy for the wealthiest companies on earth, paid by people whose electric bills have nothing to do with AI inference workloads. The new tariff structures are an attempt to price that risk correctly.
Existing Plants Are the Bridge — Whether Anyone Planned for It
Meanwhile, the question of where the power actually comes from is getting its own answer, and it's not the one the Ratepayer Protection Pledge crowd wanted. Constellation Energy CEO Joseph Dominguez said it plainly on an August 7 earnings call: "We're never going to build this economy if we have to wait for new power plants to be built before we can connect any data center. The bedrock of building out at least this early phase of the data economy is going to rely heavily, in my view, on existing generation."
Dominguez's framing is worth sitting with. Constellation owns roughly 22 GW of nuclear capacity — fixed-price, zero-carbon, dispatchable around the clock. His argument is that the grid has stranded capacity in non-peak hours that can serve data centers right now, and that the peak reliability problem is manageable with batteries, demand response, and peaking resources. "We have a peak capacity concern, not an energy concern," he said.
That's a civilizationally important distinction. The abundance is already there, sitting idle most hours of the year. The bottleneck is the handful of peak hours and the pricing mechanisms that haven't caught up to a world where a single customer can represent gigawatts of new load overnight.
When the Load Spikes 50%, the Equipment Wasn't Designed for That
There's a hardware dimension to this that's getting less attention than it deserves. Bloomberg reported on August 6 that AI workloads are causing power demand to spike as much as 50% above design capacity inside data centers themselves — fast enough to damage batteries, generators, and cooling systems, wearing them out far sooner than expected. Even a few minutes of lost uptime hits revenue hard.
This is the volatile demand problem landing on the physical infrastructure. AI inference isn't a steady load. It's bursty, unpredictable, and increasingly intense. The data center was designed for a different computational era. The grid interconnection was sized for a different load profile. Both are now absorbing stress they weren't engineered to handle.
The pricing implication is direct: if the load is volatile and damaging, the cost of serving it is higher than a flat-rate PPA would suggest. Markets that price electricity on average demand are systematically undercharging for peak stress. Italy's energy regulator just fined utility A2A €5 million for manipulating power prices during the 2022 energy crisis by withholding gas plant capacity — a reminder that stressed electricity markets create manipulation incentives when pricing signals are wrong. The same dynamic applies here, just in the opposite direction: underpriced peak demand creates its own distortions.
The Pricing Model Has to Catch Up to the Load
The through-line across all of this: power purchase agreements are large enough to reshape electricity market pricing, but the pricing mechanisms themselves were built for a different world. Utilities are improvising with new tariffs. Grid operators are threatening curtailment. Equipment is failing under load profiles it wasn't designed for. And the existing nuclear fleet is quietly becoming the most valuable generation asset in the country — because it's the only resource that can deliver fixed-price, always-on power at the scale data centers actually need.
Watch for FERC's response to state-level tariff innovations in Ohio and Virginia — whether the commission moves to standardize large-load interconnection requirements nationally, or leaves states to experiment independently, will determine whether the stranded cost problem gets solved systematically or piecemeal. That decision shapes every PPA signed for the next decade.
The future is electric. The pricing model just has to get there first.
